Deep learning models for time series forecasting provide high representational capacity, making them well-suited for capturing complex temporal patterns. However, their effectiveness often hinges on large volumes of training data and significant computational resources. To mitigate these limitations, Esri has integrated Time-MoE, a pretrained time series foundation model designed for high-performance forecasting directly without requiring fine-tuning.
Time-MoE is trained on Time-300B, massive corpus comprising more than 300 billion time points across diverse domains. This extensive pretraining enables strong generalization capabilities, allowing the model to perform inference-only forecasting across a wide range of real-world scenarios without retraining. Built on a sparse Mixture-of-Experts (MoE) architecture, Time-MoE selectively activates a subset of its 2.4 billion parameters per prediction. This architectural efficiency allows it to deliver high accuracy while maintaining computational scalability. It supports flexible input context lengths up to 4,096 time steps and adaptable forecasting horizons, allowing users to forecast over both short and long term with high accuracy.
Time-MoE is optimized for univariate time series forecasting using a one-step-ahead prediction approach. Its scalability and adaptability make it ideal for large-scale forecasting applications such as regional weather prediction, energy demand forecasting, and financial market analysis. By using this model, organizations and researchers can achieve faster and more accurate forecasts while significantly reducing the data and computational resources required to train such time series models from scratch.
Model details
This model has the following characteristics:
- Input—The netCDF cube containing the variable that will be used to forecast to future time steps. This file must have an .nc file extension and must have been created using the Create Space Time Cube By Aggregating Points, Create Space Time Cube From Defined Locations, or Create Space Time Cube From Multidimensional Raster Layer tool.
- Output—The output feature class of all
locations in the space-time cube with forecasted values stored as
fields. The layer displays the forecast for the final time step and
contains pop-up charts showing the time series and forecasts for
each location.
An optional output space-time cube (.nc file) can also be created. This will contain the values of the input space-time cube with the forecasted time steps. Use the Visualize Space Time Cube in 3D tool to see all of the observed and forecasted values simultaneously.
- Compute—This workflow is compute intensive and a GPU with compute capability of 6.0 or higher is recommended.
- Applicable geographies—This model is expected to work well across the globe.
- Architecture—The Time-MoE model is a scalable decoder-only transformer model with a sparse Mixture-of-Experts (MoE) architecture.
- Training data—The model has been trained on an enormous dataset (Time-300B) that includes more than 300 billion time points across multiple domains.
Access and download the model
Download the Time-MoE pretrained model from ArcGIS Living Atlas of the World. Alternatively, access the model directly from ArcGIS Pro 2.7, or use it in ArcGIS Online with the Professional or Professional Plus user type.
To download the model, complete the following steps:
- Browse to ArcGIS Living Atlas of the World.
- Sign in with your ArcGIS Online credentials.
- Search for Time-MoE and open the item page from the search results.
- Click the Download button to download the model.
You can use the downloaded .dlpk file directly in ArcGIS Pro, or upload and use it in ArcGIS Enterprise. Additionally, you can fine-tune the pretrained model if necessary.
Release notes
The following are the release notes:
| Date | Description |
|---|---|
| June 2025 | First release of Time-MoE pretrained model. |